If you're training or fine-tuning a speech model, open-source speech datasets are where almost everyone starts, and for good reason. Corpora like LibriSpeech, Common Voice, and GigaSpeech offer thousands of freely usable hours, established benchmarks, and ready-made recipes in ESPnet, NeMo, and Hugging Face.
This list covers the 15 most useful open ASR and TTS datasets in 2026, with real numbers on size, language coverage, and licensing: the three things that actually determine whether a dataset is usable for your project. It ends with an honest assessment of where open data stops being enough, because license terms, read-speech bias, and missing consent chains bite hardest exactly when a project moves from research to product.
The 15 Best Open Speech Datasets at a Glance
| Dataset | Size | Languages | License | Best for |
|---|---|---|---|---|
| LibriSpeech | 960 h | English | CC BY 4.0 | ASR baselines, benchmarking |
| Common Voice | 20,000+ validated h | 130+ | CC0 | Multilingual ASR, accents |
| Multilingual LibriSpeech | ~50,000 h | 8 | CC BY 4.0 | Large-scale multilingual ASR |
| People's Speech | 30,000+ h | English | CC-BY / CC-BY-SA subsets | Large-scale English ASR |
| Google Speech Commands | 105k one-second clips | English | CC BY 4.0 | Keyword spotting |
| VoxPopuli | 400k h unlabeled / 1.8k h transcribed | 23 / 16 | CC0 | Pretraining, EU languages |
| TED-LIUM 3 | 452 h | English | CC BY-NC-ND 3.0 | Research ASR (non-commercial) |
| GigaSpeech | 10,000 h transcribed | English | Custom terms | Diverse-domain English ASR |
| FLEURS | ~12 h × 102 languages | 102 | CC BY 4.0 | Multilingual evaluation |
| AMI Corpus | 100 h | English | CC BY 4.0 | Meetings, diarization |
| CHiME-5/6 | ~50 h | English | Challenge terms | Far-field, noisy ASR |
| LJ Speech | 24 h | English | Public domain | Single-speaker TTS |
| VCTK | ~44 h, 110 speakers | English | CC BY 4.0 | Multi-speaker TTS |
| LibriTTS | 585 h | English | CC BY 4.0 | TTS at scale |
| AISHELL-1 | 178 h | Mandarin | Apache 2.0 | Mandarin ASR |
ASR Datasets
1. LibriSpeech. 960 hours of read English audiobooks from the LibriVox project, segmented and aligned. Still the most-cited ASR benchmark; state-of-the-art WER on test-clean sits under 2%. Its weakness is its strength: clean read speech makes clean baselines and saturated leaderboards.
2. Mozilla Common Voice. The largest crowdsourced speech corpus: volunteers read sentences in 130+ languages, with 20,000+ validated hours and growing each release. CC0 licensing makes it the least legally complicated dataset on this list. Quality and per-language volume vary enormously, though. Some languages have hundreds of hours, others under one.
3. Multilingual LibriSpeech (MLS). Meta's extension of the LibriSpeech recipe to audiobooks in eight languages (English ~44,000 hours; German, Dutch, French, Spanish, Italian, Portuguese, Polish in the hundreds to low thousands). CC BY 4.0. The go-to for scaling European-language read-speech ASR.
4. People's Speech. MLCommons' 30,000+ hour English corpus built from publicly available, transcribed audio (government proceedings, lectures, media). Released in CC-BY and CC-BY-SA subsets intended to permit commercial use. Transcript quality is variable, so treat it as weakly supervised scale, not gold-standard labels.
5. Google Speech Commands. 105,829 one-second utterances of 35 short words ("yes," "no," "stop," digits) from thousands of speakers, CC BY 4.0. The standard benchmark for keyword spotting and on-device wake-word models, not for general transcription.
6. VoxPopuli. European Parliament recordings: ~400,000 hours of unlabeled audio in 23 languages plus 1,791 hours transcribed across 16 languages, CC0. Excellent for self-supervised pretraining and under-resourced EU languages; the domain is formal political speech.
7. TED-LIUM 3. 452 hours of TED talks with aligned transcripts. Well-prepared and widely used, but CC BY-NC-ND 3.0 means no commercial use, which disqualifies it for product training despite its popularity in papers.
8. GigaSpeech. 10,000 hours of transcribed English from audiobooks, podcasts, and YouTube (33,000 hours of total audio). Multi-domain and closer to real-world speech than LibriSpeech. Access requires agreeing to its terms, and the YouTube-sourced portion carries the rights ambiguity discussed below.
9. FLEURS. ~12 hours per language across 102 languages, n-way parallel (same sentences translated), CC BY 4.0. Too small for training, but the standard multilingual evaluation set. This is what Whisper's per-language numbers are reported on.
10. AMI Meeting Corpus. 100 hours of multi-speaker meeting recordings with close-talk and far-field microphones, transcripts, and rich annotations, CC BY 4.0. The reference corpus for diarization and meeting ASR.
11. CHiME-5/6. Real dinner-party conversations recorded on distant microphone arrays, some of the hardest public ASR audio in existence (baseline WERs historically above 50%). Distributed under challenge-specific terms; check current conditions before commercial use.
TTS Datasets
12. LJ Speech. 24 hours of a single female speaker reading public-domain texts. Public domain, universally supported by TTS toolkits. It is the "hello world" of speech synthesis.
13. VCTK. ~44 hours from 110 English speakers with varied accents, CC BY 4.0. The standard corpus for multi-speaker and voice-cloning research.
14. LibriTTS. 585 hours derived from LibriSpeech but restored to 24 kHz with punctuation and speaker metadata preserved, 2,400+ speakers, CC BY 4.0. The default for training larger TTS models; LibriTTS-R offers a quality-enhanced version.
Non-English ASR
15. AISHELL-1. 178 hours of Mandarin from 400 speakers, Apache 2.0. A rare permissively licensed Mandarin corpus and the standard open benchmark for Chinese ASR (AISHELL-3 adds a TTS-oriented multi-speaker set).
The Honest Limits of Open Speech Data
Open datasets built modern speech AI, but four problems surface when you move from benchmark to product:
License ambiguity. "Available for download" is not "licensed for commercial training." TED-LIUM is explicitly non-commercial; corpora sourced from YouTube inherit unresolved questions about whether uploaders could grant training rights at all. If your legal team asks "who consented, and to what?", most open datasets have no answer. Our speech data licensing guide breaks down what to check clause by clause.
Read-speech bias. LibriSpeech, MLS, Common Voice, and LibriTTS are all read speech. Models trained on them systematically underperform on spontaneous conversation: disfluencies, overlap, fast reduced articulation. If your product listens to real people talking (calls, meetings, voice agents), read-heavy training data is the single biggest accuracy gap you'll face.
No consent chain. GDPR treats voice as personal (and potentially biometric) data. Open corpora scraped or repurposed from the web give you no documented consent from the speakers, which is increasingly a hard blocker for enterprise procurement and EU deployments.
Saturated benchmarks and coverage gaps. Sub-2% WER on test-clean tells you nothing anymore, and language coverage is brutally uneven: English has 100,000+ open hours while commercially important languages like Hindi, Mexican Spanish, and Swahili have a small fraction of that in spontaneous conversational form.
When to license commercial data instead: when you need spontaneous conversation rather than read prompts, a documented consent and license chain, guaranteed speaker demographics and audio specs, or real volume in languages the open ecosystem neglects. That's the gap our catalog exists to fill: 60 languages, 500-2,000 conversational hours each, transcribed and consented for commercial use.
Need training data?
Open datasets are a great foundation; when you need consented conversational speech at scale, we license datasets in 60 languages at $60-95 per hour. Explore the catalog or contact us for samples.